42 static inline int CeedTensorContract_Avx_Blocked(CeedTensorContract contract, CeedInt A, CeedInt B,…  in CeedTensorContract_Avx_Blocked()  argument
52   for (CeedInt a = 0; a < A; a++) {  in CeedTensorContract_Avx_Blocked()
104 static inline int CeedTensorContract_Avx_Remainder(CeedTensorContract contract, CeedInt A, CeedInt …  in CeedTensorContract_Avx_Remainder()  argument
116   for (CeedInt a = 0; a < A; a++) {  in CeedTensorContract_Avx_Remainder()
153 static inline int CeedTensorContract_Avx_Single(CeedTensorContract contract, CeedInt A, CeedInt B, …  in CeedTensorContract_Avx_Single()  argument
164   for (CeedInt a = 0; a < (A / AA) * AA; a += AA) {  in CeedTensorContract_Avx_Single()
187   const CeedInt a = (A / AA) * AA;  in CeedTensorContract_Avx_Single()
192     for (CeedInt aa = 0; aa < A - a; aa++) {  in CeedTensorContract_Avx_Single()
200         for (CeedInt aa = 0; aa < A - a; aa++) {  // unroll  in CeedTensorContract_Avx_Single()
205     for (CeedInt aa = 0; aa < A - a; aa++) {  in CeedTensorContract_Avx_Single()
210   const CeedInt A_break = A % AA ? (A / AA) * AA : (A / AA - 1) * AA;  in CeedTensorContract_Avx_Single()
245       for (CeedInt a = A_break; a < A; a++) v[a * J + j] += tq * u[a * B + b];  in CeedTensorContract_Avx_Single()
254 static int CeedTensorContract_Avx_Blocked_4_8(CeedTensorContract contract, CeedInt A, CeedInt B, Ce…  in CeedTensorContract_Avx_Blocked_4_8()  argument
256   return CeedTensorContract_Avx_Blocked(contract, A, B, C, J, t, t_mode, add, u, v, 4, 8);  in CeedTensorContract_Avx_Blocked_4_8()
258 static int CeedTensorContract_Avx_Remainder_8_8(CeedTensorContract contract, CeedInt A, CeedInt B, …  in CeedTensorContract_Avx_Remainder_8_8()  argument
260   return CeedTensorContract_Avx_Remainder(contract, A, B, C, J, t, t_mode, add, u, v, 8, 8);  in CeedTensorContract_Avx_Remainder_8_8()
262 static int CeedTensorContract_Avx_Single_4_8(CeedTensorContract contract, CeedInt A, CeedInt B, Cee…  in CeedTensorContract_Avx_Single_4_8()  argument
264   return CeedTensorContract_Avx_Single(contract, A, B, C, J, t, t_mode, add, u, v, 4, 8);  in CeedTensorContract_Avx_Single_4_8()
270 static int CeedTensorContractApply_Avx(CeedTensorContract contract, CeedInt A, CeedInt B, CeedInt C…  in CeedTensorContractApply_Avx()  argument
275     for (CeedInt q = 0; q < A * J * C; q++) v[q] = (CeedScalar)0.0;  in CeedTensorContractApply_Avx()
280     CeedTensorContract_Avx_Single_4_8(contract, A, B, C, J, t, t_mode, true, u, v);  in CeedTensorContractApply_Avx()
283 …if (C >= blk_size) CeedTensorContract_Avx_Blocked_4_8(contract, A, B, C, J, t, t_mode, true, u, v);  in CeedTensorContractApply_Avx()
285 …if (C % blk_size) CeedTensorContract_Avx_Remainder_8_8(contract, A, B, C, J, t, t_mode, true, u, v…  in CeedTensorContractApply_Avx()